Research libraryCotton to Garment Across Africa
Chapter 13 of 14 · Part IV · The GFT Synthesis · Draft web edition
A greenfield plant has no operating history, so GFT's proposed stack builds a six-family KPI ledger and validates edge inspection from the first shift, then admits forecasting and generative design only when their data and physical-validation gates are passed.
The cloth is woven from this chapter's 6 exhibits: each weft is one exhibit's own data, normalised; each warp group is a section. The probe reads the exhibit's own value where it crosses. Procedural visualisation, not a measurement.
Read-out
Illustrative GFT architecture for a greenfield factory. Local inference and manual fallback protect operations from connectivity loss. Forecasting and generative design remain gated. This is not a claim of implemented GFT infrastructure.
Unit: system responsibility. Status: illustrative. Sources: R27, R29, S1301, S1303, S1307. Illustrative figures are teaching examples, not measured factory results.
A greenfield African garment factory should not buy artificial intelligence as a substitute for an operating system. It should build the operating system first, capture trustworthy production events from the first shift, and add algorithms only where each has a defined decision, a measurable denominator and a safe failure mode. GFT's answer is a sequenced stack: a formula-locked performance-management workbook establishes the system of record; edge computer vision supports inspection without waiting for sales history; demand forecasting starts only after order data passes a maturity test; and generative design remains behind engineering and physical-sample gates. This sequence gives a new factory a credible route towards Asian conversion efficiency without pretending that it already possesses legacy data or proven operating results.
The distinction matters because the technologies do not share the same data dependency. A forecasting model learns from the factory's own orders, shipments, stockouts and returns. A pre-trained inspection model can instead begin with external visual examples, provided the factory validates it on local fabric, lighting, cameras and defect definitions. Published reviews report manual inspection success around 60 to 70 percent and automated methods above 90 percent, but those figures are research benchmarks rather than guaranteed performance in a new African plant.[R29][S1302] Likewise, McKinsey reports substantial forecasting gains across supply-chain cases, not an entitlement to the same gain in every garment business.[S1301] GFT therefore treats each external result as a hypothesis to retest at line level.
Matching an Asian benchmark is consequently a measurement protocol, not a marketing statement. The comparison must use the same garment operation, standard allowed minute basis, size mix, shift calendar, quality rule and inspection boundary. A factory can claim parity only after a sustained, reconciled period shows comparable earned minutes per attended minute, first-pass yield and delivery performance. Until then, the stack described here is GFT's proposed design for its Cameroon textile programme, not an account of an operating GFT factory or a realised efficiency result.
13.1
Read-out
Mordor Intelligence supplies the 2025 and 2030 values. Market.us supplies the 2035 value and the separate 2024 apparel share discussed in prose. Vendor definitions and forecast methods may differ. Do not splice these bars into one time series or infer a common growth rate.
Unit: US$ billion, nominal market estimates. Status: sourced. Sources: R24, R25.
Read this exhibit in its chapter ↗Read-out
GFT design informed by published process and utility categories. Layer ownership, metering and escalation are recommendations, not verified park implementation.
Unit: illustrative responsibility architecture. Status: illustrative. Sources: S04, S15. Illustrative figures are teaching examples, not measured factory results.
Read this exhibit in its chapter ↗13.1 · Introduction
The greenfield cold-start problem is not simply an absence of rows in a database.
Exhibit 02, chapter 01: Mordor 2025
13.2 · Simple baselines remain necessary when order history is thin
A new plant begins without its own reorder patterns, style-level seasonality or reliable stockout history.
Exhibit 02, chapter 01: Mordor 2025
13.3 · Edge-first computing protects the decision from the connection
The architecture also has a physical constraint.
Exhibit 18, chapter 04: Decision
13.1 · the bodyThe film above follows these sub-steps.
The greenfield cold-start problem is not simply an absence of rows in a database. It is an absence of shared meaning. A sewing-line output total is ambiguous if one shift counts pieces leaving the final operation while another counts pieces passing final inspection. Downtime is unusable if operators mix planned changeovers with breakdowns. A defect record has little analytical value when the same hole is described as fabric damage, a cutting fault and a sewing reject. Models trained on such records inherit the inconsistency and conceal it behind numerical confidence.
The first design task is therefore a data contract. Every event needs a timestamp, Line ID, Operator ID where appropriate, Lot ID, style and size identifier, operation code, quantity, unit and source device or document. Every quantity also needs a status: observed, corrected, estimated or rejected. The data dictionary must state whether output means sewn, inspected, repaired or packed, and whether efficiency uses attended time, available time or paid time. These decisions are managerial and engineering choices. They should be approved before software configuration, because a later change in definition can invalidate an entire time series.
A new plant begins without its own reorder patterns, style-level seasonality or reliable stockout history. McKinsey's review of data-light forecasting does not prescribe one universal waiting period. It recommends choosing model complexity to suit the sample, smoothing genuinely anomalous periods, preparing scenarios for uncertainty and adding external data only where relevant.[S1301] That evidence supports a more disciplined response than either premature machine learning or a blanket refusal to forecast. During early operation, planners should use visible baseline methods: contracted order books, customer-confirmed delivery windows, moving averages, simple seasonal assumptions and explicit scenarios. Each assumption must remain editable by the responsible planner.
External data cannot repair a weak internal boundary. Weather may help a maintenance forecast, and a customer's point-of-sale feed may help a replenishment forecast, but neither reveals what a factory failed to ship because its own line stopped. Orders, shipments and true demand must remain separate fields. A stockout is a censored observation of demand, not proof that demand equalled dispatched quantity. Preserving these distinctions from the first order prevents a later forecasting team from learning the wrong target.
The architecture also has a physical constraint. Inspection and line intervention cannot wait for a remote service to respond. NIST identifies decentralised applications, management and analytics as a response to cloud environments constrained by scale, heterogeneity and latency.[S1303] In GFT's proposed design, inference runs beside the inspection frame or quality table. The edge device retains the model, recent event buffer and local rules. It writes a compact result to the factory ledger when connectivity permits. Production continues if the external network fails.
Edge-first does not mean isolated or uncontrolled. Devices need authenticated administration, signed model versions, restricted ports, backed-up configurations and a tested manual fallback. The vision system should recommend a disposition, not actuate a cutter, sewing machine or safety device. A quality technician retains authority over acceptance and rework. This separation limits the consequence of a false classification and keeps functional safety outside the AI design. NIST's voluntary AI Risk Management Framework similarly places trustworthiness across design, development, use and evaluation rather than treating deployment as the end of control.[S1307]
Exhibit 59 sets out the proposed responsibility boundaries. It is an architecture to implement and test, not evidence that the layers already operate together.
The exhibit's important feature is not the number of layers. It is the direction of authority. Production generates events; the signal layer validates them; control determines what may happen; learning offers recommendations within that mandate; and named managers remain accountable for action. Data can move upwards for later analysis, but an unvalidated model never acquires direct authority over machinery.
13.4
GFT's proposed scorecard contains 24 core measures across six families. The count is less important than the chain from a physical event to a management decision. Production shows whether assets and labour create earned output. Quality shows whether that output is acceptable without repair. Delivery tests the promise to the customer. Cost converts physical losses into financial consequences. People records the capability and safety conditions behind performance. Sustainability applies the same denominator discipline to energy, water and waste. Cameroon's working-time rules and the programme's inspection evidence inform the operating context, but the formulas and thresholds below are proposed GFT controls rather than external industry standards.[R16][R28][R29]
Every measure needs one formula and one accountable owner. Overall Equipment Effectiveness (OEE) multiplies availability, performance and quality for a defined asset and period; it must not be presented as sewing-line labour efficiency. Standard Allowed Minute (SAM) efficiency divides earned standard minutes by attended direct-labour minutes. First-Pass Yield (FPY) counts units accepted without rework divided by units entering inspection. Defects per Hundred Units (DHU) counts defects rather than defective garments, so one garment can contribute more than one defect. On-Time In-Full (OTIF) is binary at order-line level unless the customer contract specifies another rule.
Exhibit 60 compresses the scorecard into its operating contract. The formulas should sit beside the source fields in the workbook so that a reviewer can reproduce every reported result.
Read-out
| Production | OEE = A x P x Q; SAM efficiency = earned SAM / attended minutes; plan attainment = actual / plan; throughput = accepted units / period | Production manager | Shift and day |
|---|---|---|---|
| Quality | FPY = first-pass accepted / inspected; DHU = 100 x defects / inspected; rework share; supplier fault share | Quality manager | Shift and week |
| Delivery | OTIF = complete on-time order lines / due lines; order cycle = dispatch minus receipt; production-to-ship lag; supplier OTIF | Planning lead | Day and week |
| Cost | Cost / earned SAM; marker = pattern / marker area; trim cost variance | Finance and IE leads | Week and month |
| People | Attendance; turnover; training hours / worker; LTIFR = injuries x exposure factor / hours worked | HR and safety leads | Week and month |
| Sustainability | kWh / accepted garment; water litres / kg where applicable; segregated waste share; landfill intensity; renewable share | Utilities and EHS leads | Day and month |
GFT scorecard design. The 24 measures and their operating definitions are proposed controls, not measured factory results or universal thresholds. R16 informs working-time context; R28 and R29 inform the quality-measurement context.
Unit: proposed KPI definitions. Status: illustrative. Sources: R16, R28, R29. Illustrative figures are teaching examples, not measured factory results.
Thresholds should trigger investigation rather than automatic punishment. A red FPY result sends the quality lead to the defect Pareto and the affected lot. A low marker-utilisation result sends industrial engineering to the marker, width and size-assortment records. A high SAM efficiency accompanied by rising DHU is not a success; it may indicate that inspection has weakened or that operators are passing defects downstream. The daily review must therefore read production, quality and delivery together.
Exhibit 61 displays only that arithmetic. Its value is to make the denominator visible, not to establish an Asian benchmark or a GFT target.
Read-out
Constructed case: 25 operators x (480 - 60) minutes = 10,500 attended minutes; 12 SAM per garment. At 80% efficiency: 700 garments. At 90%: floor(10,500 x 0.90 / 12) = 787 garments. No quality or downtime change is modelled.
Unit: accepted garments per shift. Status: illustrative. Illustrative figures are teaching examples, not measured factory results.
The same logic applies to automation. A device can increase accepted output only if it improves a binding loss without creating another. Faster inspection has little value when the line already waits for fabric, and a small FPY improvement can matter more than a large speed increase when rework consumes scarce operator minutes. The workbook lets managers locate that constraint before they approve capital.
13.6
Related · chapter 037 paragraphs step it
Read-out
Illustrative GFT architecture. AbTF supports the HIP chain, now called Segregation, from bale through textile product; the garment data carrier and factory exception controls are proposed extensions. This is not a claim of implementation at a GFT operating factory.
Unit: GFT reference control sequence. Status: illustrative. Sources: S0301, S0305, S10, S34. Illustrative figures are teaching examples, not measured factory results.
Read this exhibit in its chapter ↗Metrics do not govern a factory because they exist in a file. They govern it when the same people review the same reconciled values at a fixed time and leave with an owner and deadline. GFT's illustrative management rhythm starts with a 15-minute shift huddle at each cutting and sewing line. The team reviews the previous shift's plan attainment, FPY, top defect, downtime cause and safety exception. It agrees no more than a few actions, each with a named owner. A weekly cross-functional review then examines recurring losses, supplier faults, corrective and preventive actions (CAPA), delivery risk and training needs.
The monthly Sales and Operations Planning (S&OP) meeting serves a different purpose. It does not rerun the daily defect review. It reconciles the demand plan, confirmed orders, capacity, material arrivals, inventory, cash implications and workforce plan. Executives sign off one feasible operating plan and record the assumptions that would change it. Daily huddles protect flow; weekly reviews remove repeated causes; monthly S&OP resolves trade-offs between sales and capacity. Mixing those agendas weakens all three.
For the early operating phase, GFT proposes a formula-locked Excel workbook pack as the principal system of record. This is a controlled application, not a spreadsheet emailed between personal computers. A restricted master copy sits on managed storage with role-based editing, scheduled backup, version history and a documented restore test. Input cells accept only defined types and ranges. Formula cells remain locked. A change log records who altered a master definition, when, why and with whose approval.
The relational structure matters more than the visual dashboard. Separate tables hold operators, lines, styles, operations, customers, suppliers, lots and defect codes. Transaction tables refer to those stable identifiers rather than retyping names. A Line ID must not change when a supervisor changes, and a defect code must retain its meaning across shifts. QR travellers can reduce manual transcription by carrying Lot ID and bundle identity from cutting to sewing and inspection. Manual entries still require reason codes for corrections; deleting a rejected record would destroy the audit trail.
Daily controls should reject duplicate event IDs, impossible timestamps, negative quantities, unknown defect codes and output that exceeds available input without an explained adjustment. Reconciliation then ties accepted production to inspection and packing, ties fabric issued to cut panels plus recorded waste, and ties attended minutes to approved shift records. Exceptions enter a clearance queue rather than being silently corrected. These checks provide the training data quality that later machine learning will require.
Excel is a sequencing choice, not a permanent architecture. The factory should migrate when concurrent users, transaction volume, customer interfaces or control requirements exceed the workbook's tested limits. Migration becomes safer when tables already have stable keys, documented formulas and complete histories. Conversely, buying a manufacturing execution system before those decisions are settled merely automates ambiguity. The phase gate is operational readiness, not embarrassment at using a workbook.
Interlude · the cloth half-woven
Woven: sections 13.0 to 13.6Still to weave: 13.8 to 13.15
Read-out
AbTF reports 2025 farm and production coverage in its annual report. The partner count is reported as of 19 February 2026 and concerns partners able to trace CmiA cotton, not all CmiA supply-chain participants.
Unit: reported status for 2025 and February 2026. Status: sourced. Sources: S08, S10.
Read this exhibit in its chapter ↗13.8
Read-out · sourced
Values summarise separate inspection-literature findings. They are not a specification for one machine, not measured at a GFT factory, and not equivalent to FPY. Speed and defect-class coverage must be validated for the selected fabric and inspection boundary.
Unit: reported literature benchmark. Status: sourced. Sources: R28, R29, S1302.
Computer vision occupies a different position from demand forecasting because its initial training data need not come from the new factory's order book. A supplier can provide a pre-trained model, cameras and an edge inference device. The factory still cannot accept the supplier's accuracy claim as its own operating result. Fabric construction, colour, surface texture, illumination, camera distance, line speed and defect taxonomy all alter what the model sees. Local validation is therefore part of commissioning, not a later improvement.
The external case for investigation is credible. The inherited inspection literature reports manual defect-detection accuracy of 60 to 70 percent, automated performance above 90 percent, coverage of more than 40 defect types and line-scan inspection speeds up to 60 metres a minute.[R28][R29] A 2025 review of YOLO-based fabric inspection likewise reports automated success above 90 percent against 60 to 70 percent for human inspection.[S1302] These statements do not prove that one system achieves every figure on one fabric, nor do they establish a garment factory's FPY. Detection success, classification accuracy, defect recall and first-pass yield measure different things.
Exhibit 62 keeps those external benchmarks visible while stating their limits.
Commissioning should begin in shadow mode. The camera records its proposed classification while trained inspectors continue the existing inspection and disposition. The trial set must include the actual fabric families, colours, finishes and defect types expected in production, with hard cases retained rather than removed. Quality staff label the reference result independently, and disagreements enter adjudication. A randomly selected mixture dominated by good fabric can make overall accuracy look reassuring while hiding poor detection of the faults that matter.
The release review should therefore examine a confusion matrix by defect class, including false accepts, false rejects and unclassified items. False accepts allow defects downstream; false rejects consume inspection time and may waste good fabric. Their costs are not equal across products. A small shade variation may be critical on a uniform contract but acceptable on a different product only if the customer specification permits it. The quality manager must set release thresholds against those consequences, record them in the model card and keep a manual inspection route available.
The validated model can then support, rather than replace, disposition. Each event should contain Lot ID, roll or bundle ID, timestamp, image reference, model version, predicted class, confidence score where meaningful, operator decision and final disposition. Images need a retention policy based on quality investigation and customer requirements, while cameras should be framed on material rather than workers. Periodic challenge samples and routine cleaning, focus and lighting checks detect drift. A model update returns the system to shadow mode until the affected classes pass again.
The investment case starts with the cost of the current defect escape and false-reject path. Industrial engineering should trace where defects are first created, where they are detected and how many attended minutes, fabric units and delivery delays occur between those points. Inspection at the correct earlier boundary can prevent a faulty panel from consuming sewing and finishing time. It cannot repair the process cause by itself. The resulting CAPA must return to the supplier, spreading table, needle, operator method or machine setting responsible for the recurrence.
This is why GFT places vision in Tier 1A for early deployment but does not call it autonomous quality management.[R28][R29] The measurable benefits are fewer defect escapes, less avoidable rework, faster containment and better-coded quality data. A supplier demonstration establishes technical possibility. Only a reconciled before-and-after trial on the factory's own products establishes operating value.
13.11
GFT classifies machine-learning demand forecasting as strategically important but activates it only after the underlying records pass a maturity gate. The outline's proposed observation window of 6 to 12 months is an illustrative planning rule, not a statistical law. A stable repeat uniform style may support a simple model earlier than a seasonal fashion range. A full annual cycle may still be inadequate for a new customer with no reorder history. Eligibility must be assessed by stock-keeping unit and customer segment rather than switched on for the whole factory on one date.
The outline’s proposed observation window of 6 to 12 months is an illustrative planning rule
illustrative
The evidence supports both ambition and restraint. The inherited fashion-AI register records potential inventory-cost reductions of 35 to 40 percent and forecast-error reductions of 20 to 50 percent across broader use cases.[R27] McKinsey's supply-chain-specific article reports forecast-error reductions of 20 to 50 percent, lost-sales and product-unavailability reductions up to 65 percent, warehousing-cost reductions of 5 to 10 percent and administration-cost reductions of 25 to 40 percent.[S1301] These sources cover different decisions and denominators. Their ranges must not be combined, transferred directly into a greenfield investment case or described as GFT outcomes.
The first candidate model competes against the current planning method, not against an imaginary perfect forecast. For each eligible style and customer, the team freezes a time-ordered training period and a later test period. It compares the candidate with a naive repeat, moving average or current planner baseline using weighted absolute percentage error where volumes permit, mean absolute error and forecast bias. The review examines error at the actual decision level, such as weekly fabric requirement or monthly style-colour demand, rather than celebrating an aggregate that hides product-level failure.
Data acceptance comes before model acceptance. Required fields include order receipt, requested and confirmed delivery dates, cancellations, shipments, returns, stockouts, price changes and customer promotions where applicable. The team must identify periods when capacity constrained sales and distinguish them from weak demand. A forecast that learns from shipments alone may reproduce past production constraints. S&OP must also approve overrides and record their reasons so that subsequent review can determine whether the model or the human adjustment added value.
Deployment begins as decision support. The model proposes a range and the planner accepts, changes or rejects it with a reason code. Procurement and capacity commitments remain inside the signed S&OP process. Drift monitoring tracks errors by segment and triggers reversion to the baseline when performance deteriorates. The business case recognises value only after reconciled changes in inventory, service and planning effort, with no double counting between those effects.
Exhibit 63 summarises the proposed sequence. Dates are observation windows, not automatic permission to deploy.
Illustrative sequencing model. Every phase requires its data, validation and owner gates. The 6 to 12 month forecasting eligibility window is a GFT planning rule, not a statistical minimum established by the cited sources. Tier 3 remains research watch.
Read-out
Illustrative sequencing model. Every phase requires its data, validation and owner gates. The 6 to 12 month forecasting eligibility window is a GFT planning rule, not a statistical minimum established by the cited sources. Tier 3 remains research watch.
Unit: months from first commercial shift. Status: illustrative. Sources: R27, R29, R30, R31, S1301, S1306. Illustrative figures are teaching examples, not measured factory results.
The roadmap makes two forms of progress compatible. The factory can use externally trained vision immediately after local validation while its own commercial history accumulates. At the same time, simple forecasting and scenario planning remain active until a more complex candidate proves superior. Delay is therefore not inactivity; it is the period in which the factory creates the evidence required for safe automation.
13.13
Multi-echelon inventory optimisation is useful only when the factory can distinguish protective stock from concealed instability. Imported trims, regional fabric, work in progress and finished uniforms have different lead times, demand patterns and consequences of shortage. The illustrative Tier 2 pilot therefore starts with a small set of items whose demand, supplier lead time, minimum order quantity and service requirement are complete. It compares current reorder rules with a candidate policy at each stocking point. Stable OEE and FPY are prerequisites because unpredictable rework and stoppage otherwise enter the model as false demand variability.
The pilot must preserve physical and financial reconciliation. System stock should agree with a controlled count; open purchase orders need confirmed rather than hoped-for dates; and fabric reserved for one contract must not be treated as available to another. Candidate rules are tested on a historical period and then in a bounded live trial. Service, inventory value, expiry or obsolescence, emergency freight and planner workload are assessed together. An algorithm that reduces average stock by creating shortages on a critical trim has failed even if its aggregate working-capital score improves.
Short-run digital textile printing follows the same evidence discipline. It can place logos, badges and capsule designs close to dispatch without engraving a new screen for every design, while the appropriate ink and fabric process still requires testing. FESPA cites an industry estimate that digital textile printing can save as much as 95 percent of industrial water and reduce energy consumption by 75 percent against analogue alternatives.[S1304] GFT treats these as trade-literature upper claims, not base-case savings. The factory must meter water and energy across pre-treatment, printing, curing, washing and cleaning for its own process boundary.
…can save as much as 95 percent of industrial water and reduce energy consumption…
S1304
A printing pilot should therefore compare total accepted output, changeover time, colour fastness, hand feel, chemical compliance, ink consumption, curing energy, water, rejects and unit contribution. On-demand customisation may reduce finished-stock risk, but it does not remove the need for quality and effluent controls. Scale follows a passed customer specification and a reconciled unit-economics result, not the promise printed in an equipment brochure.
13.14
Human-reviewed design concept
Human-reviewed design concept. Starting point. Leads to Structured and graded pattern (Engineer, do not auto-export). · illustrative
Generative design has a different maturity profile from inspection and forecasting.
Structured and graded pattern
Structured and graded pattern. From Human-reviewed design concept (Engineer, do not auto-export), Drape and seam validation (Correct failed geometry), Cut and sew physical sample (Correct failed sample). Leads to Drape and seam validation. · illustrative
Drape and seam validation
Drape and seam validation. From Structured and graded pattern. Leads to Cut and sew physical sample (Pass virtual checks), Structured and graded pattern (Correct failed geometry). · illustrative
Cut and sew physical sample
Cut and sew physical sample. From Drape and seam validation (Pass virtual checks). Leads to Pattern and marker approval (Pass physical checks), Structured and graded pattern (Correct failed sample). · illustrative
Pattern and marker approval
Pattern and marker approval. From Cut and sew physical sample (Pass physical checks). Leads to Controlled production cutting. · illustrative
Controlled production cutting
Controlled production cutting. From Pattern and marker approval. End point. · illustrative
Illustrative GFT control sequence informed by design research. Textile IR integration remains unimplemented in the cited proposal. Physical construction and pattern-engineer approval remain mandatory; no direct image-to-cutter route is permitted.
Unit: engineering validation sequence. Status: illustrative. Sources: R30, R31, S1305, S1306. Illustrative figures are teaching examples, not measured factory results.
13.14 · the bodyExhibit 64 above; its figures are illustrative where marked.
Generative design has a different maturity profile from inspection and forecasting. The inherited research register places recent apparel image generation mainly across style, silhouette, colour and pattern ideation.[R30] A comparative study of large language models found that all tested models struggled with complex fabric constructions and intricate designs from images.[S1305] These are reasons to use generated images as concepts for human review, not as cutting instructions. Visual plausibility does not establish seam length compatibility, grain direction, seam allowance, size grading, fabric drape or needle feasibility.
Research is moving towards structured representations that connect CAD, physics simulation and sustainability assessment. The Textile IR proposal explicitly combines manufacturing-valid CAD with drape and stress validation, but its authors also state that integration of the separately validated components awaits implementation.[S1306] GFT accordingly keeps this field in Tier 3 research watch. The relevant question is not whether a system can produce an attractive garment image. It is whether a pattern engineer can reproduce, edit, grade, simulate, cut and sew the output under controlled material conditions.
Exhibit 64 shows the minimum proposed route. A failed simulation or sample returns the design to structured pattern development; neither route bypasses human approval.
A serious research trial would measure more than image quality. Pattern engineers would score editability and construction correctness; the laboratory would verify material parameters; and the sample room would record sewing defects, dimensions and fit against the approved specification. The team would retain failed cases and model versions so that progress can be audited. Until those tests demonstrate repeatable value on the factory's own products, generative design belongs in a bounded research budget, not in the critical path of customer delivery.
13.15
The proposed stack can close information delays, but it cannot guarantee conversion-efficiency parity with an Asian factory. Proof requires a like-for-like comparison across a sustained operating period. Industrial engineering must align the style operation breakdown and SAM, then normalise attended time, shift calendars, size mix, fabric behaviour and customer quality rules. Reports should show distributions and exceptions, not only averages that conceal a difficult product group.
The comparison should read SAM efficiency beside FPY, DHU, OTIF, conversion cost and energy per accepted garment. Higher output achieved by accepting weaker quality or late delivery is not parity. Financial comparisons also need a stated currency and price date, while external benchmarks need their original denominator and period. Where those conditions are unavailable, GFT should report an internal trend and withhold the cross-country claim.
This is the practical role of Warp and Signal. Production creates the garment; measurement creates the evidence that improves the next one. The workbook preserves that evidence, edge inference shortens selected inspection decisions, and later models earn authority through out-of-sample performance. The sequence gives a greenfield factory a disciplined way to learn from its own operation without waiting for a legacy system or pretending that purchased intelligence is inherited competence.
Check the record
Every knot is a source this chapter cites; select one to read its record, or filter the register by grade. Publication dates and inherited-source limits stay explicit. Read the paper's full evidence summary.
13 sources · 33 citation marks on this pageMost cited: R29, 5 times
Grade AGrade BGrade CKnot size: citations on this page
All 13 records shown
R16 · Grade C · inherited record
Inherited evidence R16
Informal share of employment about 90 percent; standard 40-hour week, 48-hour maximum, overtime at 1.2 times.
Publication date unrecorded. Inherited source-text record. The original publication has not been independently reopened for this web edition.
R27 · Grade B · inherited record
Inherited evidence R27
AI at scale in fashion: inventory cost down 35 to 40 percent; forecast accuracy up 20 to 30 percent; revenue per customer up 15 to 25 percent; forecast errors cut 20 to 50 percent; lost sales down up to 65 percent; 75 percent of executives prioritising; sector operating-profit potential US$150 to 275 billion over three to five years.
Publication date unrecorded. Inherited source-text record. The original publication has not been independently reopened for this web edition.
R28 · Grade C · inherited record
Inherited evidence R28
Fabric-defect-detection AI market US$523.4 million (2024) to US$6,522.1 million (2033) at 33.8 percent a year; Asia-Pacific about 46 percent in 2024.
Publication date unrecorded. Inherited source-text record. The original publication has not been independently reopened for this web edition.
R29 · Grade C · inherited record
Inherited evidence R29
Manual defect inspection 60 to 70 percent accurate, missing 20 to 30 percent of defects; computer vision above 90 percent across more than 40 defect types; line-scan inspection up to 60 metres a minute.
Publication date unrecorded. Inherited source-text record. The original publication has not been independently reopened for this web edition.
R30 · Grade C · inherited record
Inherited evidence R30
Generative AI in apparel 2023-2024 mainly design-side image generation (Midjourney, Stable Diffusion) across style, silhouette, colour and pattern.
Publication date unrecorded. Inherited source-text record. The original publication has not been independently reopened for this web edition.
R31 · Grade C · inherited record
Inherited evidence R31
By 2025-2026 the field moved to multi-model garment-design pipelines (fabric construction, shade, weave, silhouette) and physics-constrained generation coupling neural generation with simulation validation.
Publication date unrecorded. Inherited source-text record. The original publication has not been independently reopened for this web edition.
c13-S1301 · Grade B
AI-driven operations forecasting in data-light environments
Applying AI-driven forecasting to supply chain management, for example, can reduce errors by between 20 and 50 percent, and translate into a reduction in lost sales and product unavailability of up to 65 percent.
2022-02-15. Register access date: 2026-10-30. Source provenance is supplied by the paper evidence register.
Open the recorded source ↗c13-S1302 · Grade B
YOLO Object Detection for Real-Time Fabric Defect Inspection in the Textile Industry: A Review of YOLOv1 to YOLOv11
Automated inspection methods achieve a high success rate of over 90% compared to 60, 70% for human inspection.
2025-04-03. Register access date: 2026-10-30. Source provenance is supplied by the paper evidence register.
Open the recorded source ↗c13-S1303 · Grade A
Fog Computing Conceptual Model, NIST SP 500-325
One solution is to decentralize applications, management, and data analytics into the network itself using a distributed and federated compute model.
2018-03-14. Register access date: 2026-10-30. Source provenance is supplied by the paper evidence register.
Open the recorded source ↗c13-S1304 · Grade C
The Environmental Benefits of Waterless Digital Textile Print Ink Systems
Digital textile printing can save as much as 95% of industrial water usage, whilst energy consumption can be reduced by 75% , leading to the minimum use of resources and textile waste (Advanced Market Analytics).
2021-06-28. Register access date: 2026-10-30. Source provenance is supplied by the paper evidence register.
Open the recorded source ↗c13-S1305 · Grade C
Transforming Fashion with AI: A Comparative Study of Large Language Models in Apparel Design
Overall, all the models struggled to recognize complex fabric constructions and intricate designs from images, and relying too much on AI might hinder human creativity.
2025-09-04. Register access date: 2026-10-30. Source provenance is supplied by the paper evidence register.
Open the recorded source ↗c13-S1306 · Grade C
Textile IR: A Bidirectional Intermediate Representation for Physics-Aware Fashion CAD
Computational feasibility of individual components (GarmentCode, MPMAvatar, PEF methodology) has been validated in their respective publications; their integration awaits implementation.
2026-01-06. Register access date: 2026-10-30. Source provenance is supplied by the paper evidence register.
Open the recorded source ↗c13-S1307 · Grade A
AI Risk Management Framework
The NIST AI Risk Management Framework (AI RMF) is intended for voluntary use and to improve the ability to incorporate trustworthiness considerations into the design, development, use, and evaluation of AI products, services, and systems.
2023-01-26. Register access date: 2026-10-30. Source provenance is supplied by the paper evidence register.
Open the recorded source ↗R29 MDPI; industry inspection literature. Publication date unrecorded; inherited source-text check.
S1302 MDPI, Sensors. 2025-04-03.
S1301 McKinsey & Company. 2022-02-15.
S1303 National Institute of Standards and Technology (NIST). 2018-03-14.
S1307 National Institute of Standards and Technology (NIST). 2023-01-26.
R27 McKinsey, The State of Fashion 2025. Publication date unrecorded; inherited source-text check.
R16 WageIndicator. Publication date unrecorded; inherited source-text check.
R28 DataIntelo. Publication date unrecorded; inherited source-text check.
R30 Design Journal (Taylor and Francis). Publication date unrecorded; inherited source-text check.
R31 arXiv 2509.04705. Publication date unrecorded; inherited source-text check.
S1306 arXiv. 2026-01-06.
S1304 FESPA. 2021-06-28.
S1305 arXiv. 2025-09-04.
Keep the thread · chapter 13 of 14
Move the shuttle along the rack to preview any chapter: its sections, exhibits and reading time. Draft web edition. This research is not investment, legal or tax advice; illustrative models are not measured factory results.
The paper's rack · 14 boltsBolt length: reading time
01Part I · The Continental Paradox and the Four Value-Chain Archetypes
Africa's cotton paradox is a conversion deficit, not a fibre deficit. Durable value retention requires downstream demand, partnered midstream capacity and measurement at every transfer of ownership.
02Part I · The Continental Paradox and the Four Value-Chain Archetypes
African textile industrial policies divide into raw-cotton SEZ converters, preference-driven CMT enclaves, Mediterranean fast-fashion near-shorers and domestic retail localisation compacts.
03Part I · The Continental Paradox and the Four Value-Chain Archetypes
Organised smallholders, rain-fed cultivation and independent verification can give African lint a valuable origin record, but only segregated lots, measured fibre quality and local conversion turn that record into durable value.
04Regional archetypes in practice
Benin and Togo show how shared infrastructure can bring cotton conversion stages together. Published capacities and investment announcements establish a direction, while operating records must establish the realised result.
05Part II · Regional Archetypes in Practice: West, Central, East, Southern and North Africa
Cameroon's 95 to 5 processing gap, CICAM's operational decline and the US$300 million Camtext SA timeline create space for an agile, sustainability-branded private garment maker.[^R06][^R09][^R11]
06Part II · Regional Archetypes in Practice: West, Central, East, Southern and North Africa
Ethiopia's Hawassa Industrial Park and Kenya's Athi River and Mombasa Export Processing Zones demonstrate both the rapid job-creation power of the AGOA Special Rule and the structural fragility of single-market assembly.
07Part II · Regional Archetypes in Practice: West, Central, East, Southern and North Africa
Madagascar and Mauritius hedge US exposure through European Economic Partnership Agreements and regional yarn integration, while South Africa's R-CTFL Master Plan 2030 anchors factories on domestic retail chains.[^S13][^S17][^S26]
08Part II · Regional Archetypes in Practice: West, Central, East, Southern and North Africa
Egypt and Morocco employ over 1.24 million textile and apparel workers and ship within days to Europe and the United States, yet still import billions of dollars of yarn and fabric annually because domestic spinning and textile recycling lag garment assembly.[^S19][^S22][^S23][^S0801]
09Part III · Cross-Cutting Determinants of Competitiveness Across African Corridors
In African textiles and apparel, tariff schedules, origin rules and used-clothing border enforcement outweigh nominal wage gaps; bankable projects anchor on regional customs unions and bilateral Economic Partnership Agreements rather than fragile unilateral waivers.
10Part III · Cross-Cutting Determinants of Competitiveness Across African Corridors
A low monthly wage becomes a low cost per garment only when standard minutes, saleable output and supervisory capability are engineered together. GFT therefore treats operator and technician training as production capital, not overhead.
11Part III · Cross-Cutting Determinants of Competitiveness Across African Corridors
Factory siting is a utility decision before it is a geography decision: dry garment assembly favours the Douala-Dibamba corridor, while northern solar and cotton proximity cannot compensate for unproven power, water, effluent and logistics conditions.
12Part III · Cross-Cutting Determinants of Competitiveness Across African Corridors
European product, labour and due-diligence rules reward suppliers that can prove each material transformation. CmiA offers a useful origin baseline, but an African garment becomes inspectable only when bale, yarn, fabric, production and claims records reconcile.
13Part IV · The GFT Synthesis: Modular Entry, Partnered Midstream and the Measured Factory
A greenfield plant has no operating history, so GFT's proposed stack builds a six-family KPI ledger and validates edge inspection from the first shift, then admits forecasting and generative design only when their data and physical-validation gates are passed.
14Part IV · The GFT Synthesis: Modular Entry, Partnered Midstream and the Measured Factory
Capturing more of Africa's cotton-to-garment value requires a gated roadmap in which domestic and regional cash flows, partnered midstream mills and audited shop-floor indicators govern each capital tranche.